orbit-robotics 0.2.0

Predict whether your robot learning data will actually train successfully

ORBIT

Predict whether your robot learning data will actually train successfully.

ORBIT analyzes your robot demonstration dataset and predicts your policy's success rate — before you spend hours training. It connects to any LeRobot dataset on HuggingFace Hub.

Quick Start

pip install orbit-robotics
orbit analyze lerobot/aloha_sim_transfer_cube_human

What You Get

• Quality score with component breakdown (position diversity, action diversity, consistency, temporal coverage)
• Dead joint detection — catches stuck servos that waste training compute
• Gripper analysis — continuous and discrete gripper detection via bimodal clustering
• Directional bias detection — distinguishes goal-directed motion from data collection problems
• Calibrated success rate prediction with confidence interval, benchmarked against 82 published results
• Policy fit analysis — ACT, Diffusion Policy, SmolVLA, DP3, BC, BC-RNN
• Community comparison against 82 benchmarked datasets from published papers
• Actionable recommendations with specific joint/episode numbers

Commands

orbit analyze <dataset>                  # full analysis
orbit analyze <dataset> --json           # machine-readable output
orbit analyze <dataset> --policy act     # specific policy fit
orbit analyze <dataset> --skip-embeddings --skip-ai-assessment  # fast mode
orbit benchmark                          # browse 82 benchmark entries
orbit benchmark --task pick_and_place --min-success 0.7
orbit benchmark aloha --top 5
orbit plan "pick up cups" --robot so100 --policy act

Policy Support

Policy	Flag	Notes
ACT	--policy act	Action Chunking Transformer — needs consistent, high-res demos
Diffusion Policy	--policy diffusion_policy	Handles multimodal data well
SmolVLA	--policy smolvla	Vision-Language-Action — language-conditioned
DP3	--policy dp3	3D Diffusion Policy
BC	--policy bc	Behavioral Cloning baseline
BC-RNN	--policy bc_rnn	Recurrent Behavioral Cloning

Use --policy auto (default) to let ORBIT recommend the best policy for your data.

Robot Support

Robot	Type	Arms
SO-100	Desktop arm	1
SO-101	Desktop arm	1
Koch v1.1	Desktop arm	1
ALOHA	Bimanual	2
xArm	Industrial	1
Custom	Any	--robot custom

Advanced Usage

VLM-enhanced analysis

pip install orbit-robotics[vlm]
export GOOGLE_API_KEY=your_key
orbit analyze lerobot/aloha_sim_transfer_cube_human

Gemini Flash analyzes your observation frames to identify task type, failure modes, and difficulty — improving prediction accuracy.

Embedding analysis

pip install orbit-robotics[vision]
orbit analyze lerobot/pusht

SigLIP embeddings measure visual diversity across episodes and detect outliers.

JSON output

orbit analyze lerobot/pusht --json

Policy comparison

orbit analyze lerobot/pusht --policy act
orbit analyze lerobot/pusht --policy diffusion_policy

How It Works

ORBIT fetches dataset metadata and episode samples from HuggingFace Hub without downloading the full dataset. It runs signal diagnostics on every joint dimension to detect dead joints, clipping, and directional bias. Task complexity is estimated from action dimensionality, temporal structure, and coordination patterns. Policy fit scores how well your data matches the requirements of your chosen policy (episode count, consistency, action dimensions). All factors feed into a calibrated predictor benchmarked against 82 ground truth results from published papers (ACT, Diffusion Policy, BC variants across Push-T, ALOHA, RoboMimic, and more).

Install Options

pip install orbit-robotics          # Core analysis (no GPU needed)
pip install orbit-robotics[vision]  # + SigLIP embedding analysis
pip install orbit-robotics[vlm]     # + Gemini VLM task analysis
pip install orbit-robotics[all]     # Everything

Citation

@software{orbit2026,
  title  = {ORBIT: Predict Robot Policy Success from Training Data},
  author = {Lasne, Rahil},
  year   = {2026},
  url    = {https://github.com/Rahillasne/orbit-robotics}
}

License

MIT — see LICENSE for details.